Introduction – Why AI?

We have spent the last year at Moonsong Labs doing a deep dive into AI and creating an AI engineering team. This journey has helped us build out our knowledge of LLMs, inference optimizations, model fine tuning, training, rag, benchmarking, and other key technologies that are part of modern AI engineering. Historically our focus has been on pure Web3 infrastructure protocol development, but we are increasingly specializing in building at the intersection of AI and Web3. Our goal is to be leading builders at the edge of these 2 disruptive technologies, to be AI experts for Web3 builders, and to be Web3 experts for AI builders.  As we specialize, we continue to see our role as engineers, finding and implementing practical solutions to real world problems using AI and Web3. Our mission is to leverage these technologies to address real-world challenges, unlocking innovative use cases that not only have market demand but also enhance or reinvent existing solutions. This blog lays out some of my thoughts on the AI x Web3 space, and what principles we are following as we build here.

AI is the most important technology paradigm shift happening in the world today. I have no doubt that it will affect every part of our lives and I expect AI powered features, assistants, and agents to be ubiquitous and pervasive in most industries within the next 10 years.  Jensen Huang, the CEO of Nvidia says, “We are at the beginning of a new Industrial Revolution,” and I like this quote because it captures the profound impact that a technology like AI can have on the world.  It makes me think of an artisan visiting an early factory in the north of England in the late 18th century who saw an industrial loom for the first time, and I wonder what they thought.  Could they have seen then how that technology would change everything? I imagine it was very difficult to see. And it makes me wonder if we aren’t like that artisan now, seeing the seeds of change, but barely being able to imagine what the future AI powered factories of the mind will look like.

Why AI x Web3?

The intersection between AI and Web3 has drawn a lot of interest in the last year. However, the field is still nascent with most efforts being explorations to find product market fit.

To answer the question “Why AI x Web3?” I want to break down my answer into 2 parts and answer each of them in turn: How does AI help Web3?  And how does Web3 help AI? Thinking about and answering these questions has helped us refine the areas and use cases where Moonsong Labs wants to focus and spend our energy building, and the type of projects we want to support.

How does AI help Web3?

Vitalik offers a helpful breakdown of the ways AI can help Web3 is his blog “The promise and challenges of crypto + AI applications.” In the blog he discusses 4 different ways that AI helps Web3 and I think it’s useful to briefly walk through each of them.

• AI as a player in the game. This refers to AIs being participants in Web3 protocols alongside humans, and being used under human direction. AI used in this way serves as an extension of human will, and makes protocols and markets more efficient. Vitalik points out that this use follows the well established history of people using technology to gain some edge or reward in inefficient markets, like arbitrage opportunities between 2 different blockchains.
• AI as an interface to the game. The idea here is to use AIs to allow humans to more easily interact with Web3 systems. AI could help with this by interpreting human intents and translating those intents into Web3 transactions that can be executed. Web3 UX is notoriously difficult, so using AI to improve it makes a lot of sense.
• AI as the rules of the game. This use case involves incorporating AI into on-chain transactions. For example you could imagine a DAO which calls out to an AI to make key decisions – think in general of a smart contract that calls out to an AI to determine what to do. This has been a focus for several high profile projects like Ritual and others.
• AI as the objective of the game. The core insight is that if you could build an AI that inherited blockchain properties like transparency and verifiability, that AI could be trusted in a way a company controlled version never could. The vision is that building an entire decentralized AI tech stack could enable the creation of trust minimized or self-sovereign AI applications. Out of all of these I find this idea the most compelling, but not all AI use cases lend themselves to decentralized approaches, often due to performance reasons.

While we are interested in exploring all of these areas, our initial research efforts have focused on AI as an interface to the game. Coming from a Web2 background as I do, I believe that UX in Web3 is a major inhibitor to the growth of our space. Web3 products are simply too hard for regular people to use, and require a lot of knowledge of how the underlying infrastructure works. We will soon publish learnings and demos from our Unravel initiative, which uses AI to interpret end user intents based on underlying transaction logs and smart contract source code. Using AI to solve Web3’s UX challenges could be a big unlock, and a future venture studio project. 

How does Web3 help AI?

In addition to considering how AI can help Web3, we also have been thinking about the opposite – How Web3 can help AI. In many ways this is a more interesting question, simply because the market for AI developers is so much larger than the market for Web3 developers. TAM is one of the first things to weigh when deciding on projects to work on.

Web3 based initiatives have an interesting alternative approach to offer AI devs and users. For some AI use cases a centralized approach is still best and attempts to use a decentralized approach are unlikely to get any significant traction. However, there are use cases where a decentralized approach can offer things that are very hard to offer with a centralized service.  Creating products that have good product market fit and traction is critical for us. Targeting regular AI devs and users means that the product needs to be price performance competitive with any centralized offerings. We have spent a lot of the last year looking at use cases where decentralized approaches can be competitive.

There are 3 main underlying properties and capabilities that Web3 can provide to AI products and services: coordination mechanisms, sovereignty and ownership, and transparency and verifiability.  Let’s consider each of these in turn.

Coordination Mechanisms

One area where Web3 can be effective, is using it as a coordination mechanism, and getting a lot of independent and self interested participants to contribute towards an aggregate greater goal. One effective pattern is the aggregation of many suppliers creating a network effect. This network effect can be part of a 2 sided marketplace (supply and demand) that can benefit from a blockchains ability to bootstrap supply, and create an aggregate cost benefit on the demand side. When the supply and demand sides of a 2 sided marketplace are coordinated with a software protocol (vs a centralized company), the resulting service can be highly efficient and transparent.

It’s very difficult to bootstrap a 2 sided marketplace as a centralized company. It requires a large amount of capital and buildup of trust. A blockchain and token can be used effectively to bootstrap these kinds of marketplaces. Many DePIN scenarios fall into the 2 sided marketplace pattern. The best ones don’t only match demand and supply directly, but enjoy a network effect where the larger the supplier network, the more valuable the protocol is on the demand side.  This software coordinated 2 sided marketplace that has network effects is a key focus for us as it uses Web3 to do something it would be very hard to do without a blockchain.

Sovereignty and Ownership

Another key property that blockchains can bring to AI is sovereignty and ownership. The first wave of AI adoption has largely been driven by centralized companies like OpenAI, Anthoropic, Google, etc. This places an incredible amount of power and control with these companies, and forces users to share their data with these companies in order to use their services – a continuation of the centralizing dynamic we see with social and internet based services. Just as with other internet based services, data sharing doesn’t seem problematic until later, when it’s clear that the centralized provider is using this data to make money (as the old adage goes, if you don’t know what the product is, you are the product). A world in which there are only a small number of centralized AI companies that everyone uses will make them incredibly powerful and lead to potentially dystopian scenarios.

In our own AI experiments, we have experienced access issues, rate limiting, and service and performance degradation of the quality of responses with no explanation. Centralized AI services are black boxes where you never really know what you are going to get. Famously, Google’s Gemini AI service began generating images with a bias towards diversity that resulted in depicting figures, including Black, Native American, and Asian individuals, in historical contexts where these representations didn’t make any sense, such as America’s Founding Fathers. This was an important moment where many started to realize how alignment and tuning of AI models implicitly incorporate values into model outputs. And it begs the obvious questions: Whose values?  What values?

Sovereign AI is- AI where you own a model’s data and fully control its behavior without needing to trust a third party will be needed for many of the most important uses. Open weights models like Llama3.x family provide a great alternative to proprietary black box services run by an AI company, but you often still need to trust centralized companies to run these open source models for you. A fully decentralized AI tech stack would allow for full ownership and control of models and AI behavior in a way that isn’t possible using centralized services.

Transparency and Verifiability.

One of the best things that blockchains can provide is transparency and verifiability of transactions. If the black box of inference, compute, and data provenance could be made transparent and verifiable that would be a huge win for many use cases. Think of voting results or other public decisions that need to be unquestionably legitimate. The challenge is that putting meaningful parts of the AI workflow such as training, compute, and data on a blockchain isn’t yet viable. Blockchains today are designed to solve very different problems and can’t support AI workloads directly. Several different approaches have emerged to verify some elements of AI workflows on blockchains, such as zk ml, optimistic ml, and using crypto economic guarantees, each having their own security-to-performance tradeoffs.

Making the right tradeoffs and introducing blockchains in ML workflows to access Web3 properties might only be a practical approach in some use cases. As an example, I’ve been personally interested in the alignment of models to reflect personal and company values. If a large volume of the interactions with your customers happens through an LLM, wouldn’t you care about whether the tone and alignment of those agents matches the values of your company? The values and data used to align the models for a company could be well served living transparently on a blockchain- a place where anyone can see and verify the data and resulting behavior in question.

Moonsong Labs’s Approach to AI x Web3

One of the conclusions I’ve come to is that as things stand today, AI will drive change faster and more broadly than Web3. That is not to say that Web3 is not one of the most important technologies discovered in the last 15 years, and that it won’t drive significant change for how value and humans are coordinated. It’s that with AI, there is a more direct connection between the technology and improved company efficiency, cost reductions and better user and customer experiences, and thus it will attract more investment and adoption in the short term. With this in mind, it makes sense for us to focus on use cases where Web3 is helping AI vs the other way around. It also takes advantage of the fact that there is a much larger market of AI devs to build for, as compared to Web3 devs, and could offer ways to bring web3 to a mainstream audience, packaged inside of a useful solution. We focus on using Web3 as a means to an end, to create uniquely differentiated products that regular AI devs and customers want to use. I should note that this is much easier said than done. We have ruled out many interesting use cases due to performance, cost or complexity introduced by Web3.

Each project we work on should grow its customer base on its own merit of delivering value and solving user’s pain points, rather than the underlying technology stack it is built on. When making design tradeoffs we favor practical solutions vs theoretical ones. We are willing to make tradeoffs for performance to create competitive products that are used by real customers.

Our long term goal is to create a broader suite of decentralized AI services. We envision these services to co-exist and to be used alongside centralized AI services. Devs will use the right combination of centralized and decentralized services that make sense for their use case. We expect only a few use cases to require a full decentralized AI stack. Projects that we create will be designed to work with each other, and to integrate with existing centralized services and open source tooling. We will incorporate and leverage key Web3 properties where appropriate such as coordination mechanisms, sovereignty and ownership, and transparency and verifiability to create new and defensible solutions.

Here is a high level picture of the different layers of a modern AI technology stack:

Below are examples of opportunities for decentralized approaches in each layer: 

Application Layer

The application layer is where decentralized apps utilize AI to enhance transparency, control, and autonomy. Key Decentralized Use Cases include: 

• AI enhanced UX for Web3 services
• AI based oracles or smart contract execution
• Voting, governance, funding, and politically undesirable content (strong transparency/censorship resistance)
• Enterprise monetization of proprietary datasets (data sovereignty)

Whether employing a fully decentralized AI stack or a hybrid approach, this layer empowers apps to leverage the best of centralized or decentralized AI technologies.

Integration Layer

The integration layer orchestrates how apps interact with AI models to deliver efficient results. 

Decentralized Use Cases at the Integration layer include:

• Decentralized Agent networks and agentic tool infrastructures 
• Crowdsourced retrieval-augmented generation (RAG) and orchestration networks

Model Layer

The model layer focuses on the development, training, and deployment of AI models in a decentralized setting.

Use Cases incorporating decentralized approaches include:

• Crowdfunding for model development and revenue-sharing
• Incentive-based model marketplaces (e.g., Bittensor subnets)

This layer enables decentralized coordination and incentives, allowing developers to collectively contribute to and monetize model building.

Compute Layer

The compute layer handles the hardware-level execution of AI tasks, from training to inference, across decentralized networks.

Key Use Cases include:

• Decentralized training and fine-tuning of AI models
• Decentralized inference for scalable model execution and growing demand for inference time reasoning.

By distributing computational workloads across a network of participants, this layer ensures that AI models can scale efficiently by making efficient use of available compute resources.

Data Layer

The data layer manages the sourcing, curation, and storage of datasets for AI models.

Key Decentralized Use Cases include:

• Crowdsourced benchmarking, data labeling, and reinforcement learning from human feedback (RLHF)
• Decentralized marketplaces for training and fine-tuning datasets, ensuring privacy and data ownership

Decentralized approaches in this layer enables secure, self-sovereign data handling, allowing contributors to maintain control while sharing their data for AI model development and refinement.

Not every area is a good fit for a decentralized approach today, due to performance and other practical constraints. Further improvements in fields like cryptography and privacy preserving machine learning may be needed before practical and Web2 competitive solutions can be created.  At Moonsong Labs, we are looking at different parts of this AI stack for venture studio project opportunities. Our goal over time is to build an entire suite of decentralized AI services. We continue to believe in the power of Web3 to create novel forms of coordination, transparency, and ownership that can uniquely solve the problems we are facing.  However, we will select projects opportunistically focusing on the best (most viable) opportunities first.  

AI x Web3 Venture Studio Projects

kluster.ai

Our first venture studio project in the AI x Web3 space is kluster.ai which fits squarely into the model layer of the stack above. The key idea is to use Web3 as a coordination mechanism to aggregate GPU supply across many individual suppliers. Each supplier can’t run large AI workloads alone, but when coordinated together they can. In kluster.ai, aggregate GPU supply is dynamically allocated to support large model operations such as inference and fine tuning, accessible to devs via a simple API. This higher level of abstraction where devs are working with models via an API is notably different than existing GPU marketplaces where devs get access to raw hardware and take on the responsibility for making a collection of hardware productive.

kluster.ai targets regular AI devs vs only Web3 devs and so has a much larger total addressable market. There is strong AI dev demand for solutions like kluster.ai given the current scarcity of GPU based compute. kluster.ai occupies a part of the cost / performance curve that is hard for centralized companies to compete on, specifically batch inference with longer SLAs against large models like llama3 405b. A decentralized approach lets kluster.ai get access to and make GPU capacity productive that otherwise would be sitting idle, which in turn allows for a dev facing service that is cost competitive vs centralized approaches.

This project broadly falls under the DePIN category, which emphasizes the coordination aspect of blockchains when creating 2-sided marketplaces. Good DePIN projects don’t only coordinate 2 sides of a market. Equally important to projects of this type is an aggregation or network effect as the protocol grows. This is the problem a lot of first generation GPU marketplaces get wrong. They match devs directly with suppliers, kind of a direct copy of a Web2 business like Together or Replicate on a blockchain, but they don’t really benefit from the network growing in size. It’s very difficult to compete directly with centralized providers as a decentralized protocol. In contrast, kluster.ai’s aggregate compute grows as the network grows, and this aggregate compute can be assigned by the protocol to where the demand is. Getting this network effect right makes it much more likely for a Web3 protocol to succeed.

Other AI ideas we are exploring

We are exploring and prototyping project ideas in different parts of the decentralized AI stack shown above. One current area of research targets the data layer of the decentralized AI stack. We are exploring the use of both the coordination and sovereignty / ownership properties of a blockchain to allow data providers to contribute unique datasets that can be aggregated for inference and other AI use cases, while still preserving data privacy and ownership. Frontier models are already trained on the totality of the public information on the internet, so we believe the next bottleneck and scarce resource for AI, beyond GPU based compute, is high quality, curated data sets. A blockchain can provide compelling tradeoffs for aggregation and distribution for these datasets. We will publish more on this idea as we continue to refine it.

Conclusion

Moonsong Labs is shifting its focus to building at the intersection of AI and Web3. We see Web3 as a key enabling technology for many AI use cases and current challenges. This intersection presents an opportunity to build infrastructure that will benefit a broad set of AI developers, leveraging Web3 technology to address new challenges in the AI space.

Stay tuned for new Moonsong Labs AI x Web3 venture studio projects at different layers of the decentralized AI x Web3 stack.